Recent advances in machine learning and computer graphics provide a unique opportunity for the design and analysis of complex mechanical systems (materials and structures). Machine learning (ML) and other data-driven methodologies are disrupting our long-established perspective on predicting the behavior of physical phenomena. In parallel, modern experimental techniques (e.g., high-speed photography, rapid prototyping, 3D imaging, computer controlled automation) have forged a new culture of precision experimentation, enabling explorations of large design spaces that are yielding novel fundamental insights and generating unprecedented data in complex (highly nonlinear) structures.
However, whereas ML is becoming ubiquitous in our daily lives, making these methodologies an integral part of the research toolbox available for experimental engineering mechanics (EM) has evolved relatively slow. Specifically, the lack of recent advances in the usage of modern ML techniques to address structural mechanics problems is in contrast to other areas (e.g. computational materials science) where large data sets abound due to effective computational techniques. We speculate that the reasons for this are threefold: First, experiments in mechanics tend to yield data-sets that are substantially less voluminous than in other data-driven fields, thus hampering the application of ML. Second, the engineering mechanics community has a long tradition in analytical modeling based on differential equations and has been 'sky in breaking the ice' with ML-based techniques. Finally, complex mechanical systems tend to exhibit strong nonlinearities, due to, for example, geometry, buckling and large deformations. This makes its challenging to perform large-scale simulations using standard techniques (e.g. FEM), and calls for help from novel computational approaches, such as high-fidelity algorithms that were originally developed in the realm of computer animation, to generate datasets.
All in all, the time is ripe to explore the potential for activity at the interface of experiments, computer graphics, and machine learning.
Our workshop will be structured along two central pillars. Thrust I will address how ML/CG can assist in probing and disentangling the experimentally observed complexity of structural systems. Thrust II will seek to motivate and educate the EM community to generate appropriate data-sets that are amenable to modern ML methodologies. One important area is to use hybrids of experimental and synthetic data (e.g. produced through GC or other simulation techniques). Another intriguing area to explore in the workshop is the use of sequential learning to inform the design of new experiments and guide decision making in successive experimental campaigns.
Ultimately, our aim is to merge learned experiences from Thrusts I and II to arrive at innovative strategies to predictively understand the complex behavior of mechanical systems and use this understanding to both inform and drive engineering design. Examples of research areas that we hope to cover in our workshop include: optimization of structures undergoing large deformations, (combinatorial) design of metamaterials, prediction of patterns in crumpled sheets, and detection and prediction of events in flows of grains or foams.